Scholarlink

An explainable and efficient convolutional neural network (CNN) consists of convolutional layers, a new feature selection (FS) layer, a classifier, and a novel “patch row-column ranking map” (PRCRM). patchpatchTop-ranked common features selected by different FS methods are used to generate the feature accumulation matrix. The feature accumulation matrix is used to rank patch rows and patch columns in the PRCRM to explain the relationship among an input image, top-ranked features, top-ranked feature maps, and decisions of the CNN. Simulation results using the Alzheimer’s disease (AD) MRI preprocessed images for 4-class image classification with 6, 400 \(128\times 128\) images indicate that patch rows 6 and 11 (12.5% of 16 patch rows) and patch columns 7 and 8 (12.5% of 16 patch columns) are related to 27 distinct brain areas (48.2% of the 56 brain areas associated with AD), and 13 brain areas (28.9% of the 45 brain areas likely associated with AD). In addition, they are not associated with any brain areas not likely associated with AD. The fine-tuned ResNet50-FS model using a small number of selected features is more accurate and more memory-efficient than the traditional ResNet50 model using all features. Thus, the new patch row-column ranking method can generate useful 1D image-row-column information to interpret decisions of a CNN model. The hybrid 1D-2D-3D information in important brain areas associated with the top patches, top patch rows, and top patch columns will be used to make a robust, rational and explainable brain disease diagnosis.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Convolutional Neural Network with Feature Selection for Generating Explainable 1D Image Information for Brain Disease Diagnosis

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Luna M. Zhang

摘要

An explainable and efficient convolutional neural network (CNN) consists of convolutional layers, a new feature selection (FS) layer, a classifier, and a novel “patch row-column ranking map” (PRCRM). patchpatchTop-ranked common features selected by different FS methods are used to generate the feature accumulation matrix. The feature accumulation matrix is used to rank patch rows and patch columns in the PRCRM to explain the relationship among an input image, top-ranked features, top-ranked feature maps, and decisions of the CNN. Simulation results using the Alzheimer’s disease (AD) MRI preprocessed images for 4-class image classification with 6, 400 \(128\times 128\) images indicate that patch rows 6 and 11 (12.5% of 16 patch rows) and patch columns 7 and 8 (12.5% of 16 patch columns) are related to 27 distinct brain areas (48.2% of the 56 brain areas associated with AD), and 13 brain areas (28.9% of the 45 brain areas likely associated with AD). In addition, they are not associated with any brain areas not likely associated with AD. The fine-tuned ResNet50-FS model using a small number of selected features is more accurate and more memory-efficient than the traditional ResNet50 model using all features. Thus, the new patch row-column ranking method can generate useful 1D image-row-column information to interpret decisions of a CNN model. The hybrid 1D-2D-3D information in important brain areas associated with the top patches, top patch rows, and top patch columns will be used to make a robust, rational and explainable brain disease diagnosis.